graphkit-learn/gklearn/kernels/graph_kernel.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Mon Mar 30 11:52:47 2020

@author: ljia
"""
import numpy as np
import networkx as nx
import multiprocessing
import time
# from abc import ABC, abstractmethod
from sklearn.base import BaseEstimator # , TransformerMixin
from sklearn.utils.validation import check_is_fitted # check_X_y, check_array,
from sklearn.exceptions import NotFittedError
from gklearn.utils import normalize_gram_matrix


class GraphKernel(BaseEstimator): #, ABC):
	"""The basic graph kernel class.

	Attributes
    ----------
    _graphs : list
        Stores the input graphs on fit input data.
        Default format of the list objects is `NetworkX` graphs.
		**We don't guarantee that the input graphs remain unchanged during the
		computation.**

	References
	----------
	https://ysig.github.io/GraKeL/0.1a8/_modules/grakel/kernels/kernel.html#Kernel.
	"""

	def __init__(self, parallel=None, n_jobs=None, chunksize=None, normalize=True, verbose=2):
		"""`__init__` for `GraphKernel` object."""
		# @todo: the default settings of the parameters are different from those in the self.compute method.
# 		self._graphs = None
		self.parallel = parallel
		self.n_jobs = n_jobs
		self.chunksize = chunksize
		self.normalize = normalize
		self.verbose = verbose
# 		self._run_time = 0
# 		self._gram_matrix = None
# 		self._gram_matrix_unnorm = None


	##########################################################################
	# The following is the 1st paradigm to compute kernel matrix, which is
	# compatible with `scikit-learn`.
	# -------------------------------------------------------------------
	# Special thanks to the "GraKeL" library for providing an excellent template!
	##########################################################################


	def fit(self, X, y=None):
		"""Fit a graph dataset for a transformer.

		Parameters
		----------
		X : iterable
			DESCRIPTION.

		y : None, optional
			There is no need of a target in a transformer, yet the `scikit-learn`
			pipeline API requires this parameter.

		Returns
		-------
		object
			Returns self.

		"""
# 		self._is_tranformed = False

		# Clear any prior attributes stored on the estimator, # @todo: unless warm_start is used;
		self.clear_attributes()

# 		X = check_array(X, accept_sparse=True)

		# Validate parameters for the transformer.
		self.validate_parameters()

		# Validate the input.
		self._graphs = self.validate_input(X)

# 		self._X = X
# 		self._kernel = self._get_kernel_instance()

		# Return the transformer.
		return self


	def transform(self, X):
		"""Compute the graph kernel matrix between given and fitted data.

		Parameters
		----------
		X : TYPE
			DESCRIPTION.

		Raises
		------
		ValueError
			DESCRIPTION.

		Returns
		-------
		None.

		"""
		# Check if method "fit" had been called.
		check_is_fitted(self, '_graphs')

		# Validate the input.
		Y = self.validate_input(X)

		# Transform: compute the graph kernel matrix.
		kernel_matrix = self.compute_kernel_matrix(Y)
		self._Y = Y

		# Self transform must appear before the diagonal call on normilization.
		self._is_transformed = True
		if self.normalize:
			X_diag, Y_diag = self.diagonals()
			old_settings = np.seterr(invalid='raise') # Catch FloatingPointError: invalid value encountered in sqrt.
			try:
				kernel_matrix /= np.sqrt(np.outer(Y_diag, X_diag))
			except:
				raise
			finally:
				np.seterr(**old_settings)

		return kernel_matrix



	def fit_transform(self, X):
		"""Fit and transform: compute Gram matrix on the same data.

		Parameters
		----------
		X : list of graphs
			Input graphs.

		Returns
		-------
		gram_matrix : numpy array, shape = [len(X), len(X)]
			The Gram matrix of X.

		"""
		self.fit(X)

		# Transform: compute Gram matrix.
		gram_matrix = self.compute_kernel_matrix()

		# Normalize.
		if self.normalize:
			self._X_diag = np.diagonal(gram_matrix).copy()
			old_settings = np.seterr(invalid='raise') # Catch FloatingPointError: invalid value encountered in sqrt.
			try:
				gram_matrix /= np.sqrt(np.outer(self._X_diag, self._X_diag))
			except:
				raise
			finally:
				np.seterr(**old_settings)

		return gram_matrix


	def get_params(self):
		pass


	def set_params(self):
		pass


	def clear_attributes(self):
		if hasattr(self, '_X_diag'):
			delattr(self, '_X_diag')
		if hasattr(self, '_graphs'):
			delattr(self, '_graphs')
		if hasattr(self, '_Y'):
			delattr(self, '_Y')
		if hasattr(self, '_run_time'):
			delattr(self, '_run_time')


	def validate_parameters(self):
		"""Validate all parameters for the transformer.

		Returns
		-------
		None.

		"""
		if self.parallel is not None and self.parallel != 'imap_unordered':
			raise ValueError('Parallel mode is not set correctly.')

		if self.parallel == 'imap_unordered' and self.n_jobs is None:
			self.n_jobs = multiprocessing.cpu_count()


	def validate_input(self, X):
		"""Validate the given input and raise errors if it is invalid.

		Parameters
		----------
		X : list
			The input to check. Should be a list of graph.

		Raises
		------
		ValueError
			Raise if the input is not correct.

		Returns
		-------
		X : list
			The input. A list of graph.

		"""
		if X is None:
			raise ValueError('Please add graphs before computing.')
		elif not isinstance(X, list):
			raise ValueError('Cannot detect graphs.')
		elif len(X) == 0:
			raise ValueError('The graph list given is empty. No computation will be performed.')

		return X


	def compute_kernel_matrix(self, Y=None):
		"""Compute the kernel matrix between a given target graphs (Y) and
		the fitted graphs (X / self._graphs) or the Gram matrix for the fitted
		graphs (X / self._graphs).

		Parameters
		----------
		Y : list of graphs, optional
			The target graphs. The default is None. If None kernel is computed
			between X and itself.

		Returns
		-------
		kernel_matrix : numpy array, shape = [n_targets, n_inputs]
			The computed kernel matrix.

		"""
		if Y is None:
			# Compute Gram matrix for self._graphs (X).
			kernel_matrix = self._compute_gram_matrix()
# 			self._gram_matrix_unnorm = np.copy(self._gram_matrix)

		else:
			# Compute kernel matrix between Y and self._graphs (X).
			start_time = time.time()

			if self.parallel == 'imap_unordered':
				kernel_matrix = self._compute_kernel_matrix_imap_unordered(Y)

			elif self.parallel is None:
				kernel_matrix = self._compute_kernel_matrix_series(Y)

			self._run_time = time.time() - start_time
			if self.verbose:
				print('Kernel matrix of size (%d, %d) built in %s seconds.'
				  % (len(Y), len(self._graphs), self._run_time))

		return kernel_matrix


	def _compute_kernel_matrix_series(self, Y):
		"""Compute the kernel matrix between a given target graphs (Y) and
		the fitted graphs (X / self._graphs) without parallelization.

		Parameters
		----------
		Y : list of graphs, optional
			The target graphs.

		Returns
		-------
		kernel_matrix : numpy array, shape = [n_targets, n_inputs]
			The computed kernel matrix.

		"""
		kernel_matrix = np.zeros((len(Y), len(self._graphs)))

		for i_y, g_y in enumerate(Y):
			for i_x, g_x in enumerate(self._graphs):
				kernel_matrix[i_y, i_x] = self.pairwise_kernel(g_y, g_x)

		return kernel_matrix


	def _compute_kernel_matrix_imap_unordered(self, Y):
		"""Compute the kernel matrix between a given target graphs (Y) and
		the fitted graphs (X / self._graphs) using imap unordered parallelization.

		Parameters
		----------
		Y : list of graphs, optional
			The target graphs.

		Returns
		-------
		kernel_matrix : numpy array, shape = [n_targets, n_inputs]
			The computed kernel matrix.

		"""
		raise Exception('Parallelization for kernel matrix is not implemented.')


	def diagonals(self):
		"""Compute the kernel matrix diagonals of the fit/transformed data.

		Returns
		-------
        X_diag : numpy array
            The diagonal of the kernel matrix between the fitted data.
            This consists of each element calculated with itself.

        Y_diag : numpy array
            The diagonal of the kernel matrix, of the transform.
            This consists of each element calculated with itself.

		"""
		# Check if method "fit" had been called.
		check_is_fitted(self, ['_graphs'])

		# Check if the diagonals of X exist.
		try:
			check_is_fitted(self, ['_X_diag'])
		except NotFittedError:
			# Compute diagonals of X.
			self._X_diag = np.empty(shape=(len(self._graphs),))
			for i, x in enumerate(self._graphs):
				self._X_diag[i] = self.pairwise_kernel(x, x) # @todo: parallel?

		try:
            # If transform has happened, return both diagonals.
			check_is_fitted(self, ['_Y'])
			self._Y_diag = np.empty(shape=(len(self._Y),))
			for (i, y) in enumerate(self._Y):
				self._Y_diag[i] = self.pairwise_kernel(y, y) # @todo: parallel?

			return self._X_diag, self._Y_diag
		except NotFittedError:
            # Else just return both X_diag
			return self._X_diag


# 	@abstractmethod
	def pairwise_kernel(self, x, y):
		"""Compute pairwise kernel between two graphs.

		Parameters
		----------
		x, y : NetworkX Graph.
			Graphs bewteen which the kernel is computed.

		Returns
		-------
		kernel: float
			The computed kernel.

# 		Notes
# 		-----
# 		This method is abstract and must be implemented by a subclass.

		"""
		raise NotImplementedError('Pairwise kernel computation is not implemented!')


	##########################################################################
	# The following is the 2nd paradigm to compute kernel matrix. It is
	# simplified and not compatible with `scikit-learn`.
	##########################################################################


	def compute(self, *graphs, **kwargs):
		self.parallel = kwargs.get('parallel', 'imap_unordered')
		self.n_jobs = kwargs.get('n_jobs', multiprocessing.cpu_count())
		self.normalize = kwargs.get('normalize', True)
		self.verbose = kwargs.get('verbose', 2)
		self.validate_parameters()

		if len(graphs) == 1:
			if not isinstance(graphs[0], list):
				raise Exception('Cannot detect graphs.')
			elif len(graphs[0]) == 0:
				raise Exception('The graph list given is empty. No computation was performed.')
			else:
				self._graphs = [g.copy() for g in graphs[0]] # @todo: might be very slow.
				self._gram_matrix = self._compute_gram_matrix()
				self._gram_matrix_unnorm = np.copy(self._gram_matrix)
				if self.normalize:
					self._gram_matrix = normalize_gram_matrix(self._gram_matrix)
				return self._gram_matrix, self._run_time

		elif len(graphs) == 2:
			if self.is_graph(graphs[0]) and self.is_graph(graphs[1]):
				kernel = self._compute_single_kernel(graphs[0].copy(), graphs[1].copy())
				return kernel, self._run_time
			elif self.is_graph(graphs[0]) and isinstance(graphs[1], list):
				g1 = graphs[0].copy()
				g_list = [g.copy() for g in graphs[1]]
				kernel_list = self._compute_kernel_list(g1, g_list)
				return kernel_list, self._run_time
			elif isinstance(graphs[0], list) and self.is_graph(graphs[1]):
				g1 = graphs[1].copy()
				g_list = [g.copy() for g in graphs[0]]
				kernel_list = self._compute_kernel_list(g1, g_list)
				return kernel_list, self._run_time
			else:
				raise Exception('Cannot detect graphs.')

		elif len(graphs) == 0 and self._graphs is None:
			raise Exception('Please add graphs before computing.')

		else:
			raise Exception('Cannot detect graphs.')


	def normalize_gm(self, gram_matrix):
		import warnings
		warnings.warn('gklearn.kernels.graph_kernel.normalize_gm will be deprecated, use gklearn.utils.normalize_gram_matrix instead', DeprecationWarning)

		diag = gram_matrix.diagonal().copy()
		for i in range(len(gram_matrix)):
			for j in range(i, len(gram_matrix)):
				gram_matrix[i][j] /= np.sqrt(diag[i] * diag[j])
				gram_matrix[j][i] = gram_matrix[i][j]
		return gram_matrix


	def compute_distance_matrix(self):
		if self._gram_matrix is None:
			raise Exception('Please compute the Gram matrix before computing distance matrix.')
		dis_mat = np.empty((len(self._gram_matrix), len(self._gram_matrix)))
		for i in range(len(self._gram_matrix)):
			for j in range(i, len(self._gram_matrix)):
				dis = self._gram_matrix[i, i] + self._gram_matrix[j, j] - 2 * self._gram_matrix[i, j]
				if dis < 0:
					if dis > -1e-10:
						dis = 0
					else:
						raise ValueError('The distance is negative.')
				dis_mat[i, j] = np.sqrt(dis)
				dis_mat[j, i] = dis_mat[i, j]
		dis_max = np.max(np.max(dis_mat))
		dis_min = np.min(np.min(dis_mat[dis_mat != 0]))
		dis_mean = np.mean(np.mean(dis_mat))
		return dis_mat, dis_max, dis_min, dis_mean


	def _compute_gram_matrix(self):
		start_time = time.time()

		if self.parallel == 'imap_unordered':
			gram_matrix = self._compute_gm_imap_unordered()
		elif self.parallel is None:
			gram_matrix = self._compute_gm_series()
		else:
			raise Exception('Parallel mode is not set correctly.')

		self._run_time = time.time() - start_time
		if self.verbose:
			print('Gram matrix of size %d built in %s seconds.'
			  % (len(self._graphs), self._run_time))

		return gram_matrix


	def _compute_gm_series(self):
		pass


	def _compute_gm_imap_unordered(self):
		pass


	def _compute_kernel_list(self, g1, g_list):
		start_time = time.time()

		if self.parallel == 'imap_unordered':
			kernel_list = self._compute_kernel_list_imap_unordered(g1, g_list)
		elif self.parallel is None:
			kernel_list = self._compute_kernel_list_series(g1, g_list)
		else:
			raise Exception('Parallel mode is not set correctly.')

		self._run_time = time.time() - start_time
		if self.verbose:
			print('Graph kernel bewteen a graph and a list of %d graphs built in %s seconds.'
			  % (len(g_list), self._run_time))

		return kernel_list


	def _compute_kernel_list_series(self, g1, g_list):
		pass


	def _compute_kernel_list_imap_unordered(self, g1, g_list):
		pass


	def _compute_single_kernel(self, g1, g2):
		start_time = time.time()

		kernel = self._compute_single_kernel_series(g1, g2)

		self._run_time = time.time() - start_time
		if self.verbose:
			print('Graph kernel bewteen two graphs built in %s seconds.' % (self._run_time))

		return kernel


	def _compute_single_kernel_series(self, g1, g2):
		pass


	def is_graph(self, graph):
		if isinstance(graph, nx.Graph):
			return True
		if isinstance(graph, nx.DiGraph):
			return True
		if isinstance(graph, nx.MultiGraph):
			return True
		if isinstance(graph, nx.MultiDiGraph):
			return True
		return False


	@property
	def graphs(self):
		return self._graphs


# 	@property
# 	def parallel(self):
# 		return self.parallel


# 	@property
# 	def n_jobs(self):
# 		return self.n_jobs


# 	@property
# 	def verbose(self):
# 		return self.verbose


# 	@property
# 	def normalize(self):
# 		return self.normalize


	@property
	def run_time(self):
		return self._run_time


	@property
	def gram_matrix(self):
		return self._gram_matrix

	@gram_matrix.setter
	def gram_matrix(self, value):
		self._gram_matrix = value


	@property
	def gram_matrix_unnorm(self):
		return self._gram_matrix_unnorm

	@gram_matrix_unnorm.setter
	def gram_matrix_unnorm(self, value):
		self._gram_matrix_unnorm = value